Qualitative and Analogical modeling of cultural reasoning

1
Qualitative and Analogical modeling of cultural
reasoning

• Kenneth D. Forbus
• Emmett Tomai
• Morteza Dehghani
• Qualitative Reasoning Group
• Northwestern University

2
Our overall approach
Catalyze research by speeding encoding.Improve
results by decreasing tailorability Eventually,
practical modeling tools for analyst
decision-maker support
Produce models via analogicalgeneralization,
predictions via simulation
Story Workbench(w/MIT)
Interviews, surveys, cultural stories collected
Predicate calculusrepresentations ofstories,
explanations
Qualitative Concept Maps
3
Overview

• Key ideas
• Qualitative modeling
• Analogical reasoning and learning
• Practical natural language processing
• Modeling cultural models of food webs
• Qualitative models to capture content
• Analogical modeling to gain insights, construct
  classifier
• Modeling blame assignment (leave out, lack of
  time)
• Qualitative model of attribution theory
• Modeling moral decision-making
• Sacred versus secular values represented via
  qualitative order of magnitude representations
• Input representations via natural language

4
Qualitative Modeling

• Formalizes intuitive knowledge of systems with
  continuous aspects
• Levels of knowledge range from the person on the
  street to expert scientists and engineers
• Has been used in wide range of scientific and
  engineering modeling
• Design of mechanical, electrical, and hybrid
  systems, modeling ecosystems, modeling genetic
  regulation mechanisms
• Has been used as formalism for human mental
  models
• Cognitive modeling efforts, new educational
  systems
• Offers useful level of precision for social
  science work
• Responsibility ?Q- Coercion

5
Building Blocks for Analogical Processing

• SME models analogical matching
• Consistent with large body of psychological
  evidence
• Has been used to make novel psychological
  predictions
• Has been used in performance systems

Match

• SEQL models generalization from examples
• Used to model several learning experiments
• Used to make novel psychological predictions

Generalize

• MAC/FAC models similarity-based retrieval
• Does not require hand-indexing of descriptions
• Used to model several psychological experiments
• Has been used in a performance system

Retrieve
6
Practical NL Processing

• Most cognitive simulations have used hand-coded
  representations
• Problematic
• Tailorability
• Scaling up hard
• An alternative Semi-automatic NL processing
• Simplified English eases parsing, semantic
  interpretation

Hand translation or tagging
As a result of a dam on a river, 20 species of
fish are threatened with extinction. By opening
the dam for a month each year, you can save these
species, but 2 species downstream will become
extinct because of the changing water level.
Human Results
Simulation Results
7
New Workflow
Hand translation to simplified English
Predicate calculus versions of stimuli,
backgroundknowledge, stories
As a result of a dam on a river, 20 species of
fish are threatened with extinction. By opening
the dam for a month each year, you can save these
species, but 2 species downstream will become
extinct because of the changing water level.
Because of a dam on a river, 20 species of fish
will be extinct. You can save them by opening the
dam. The opening would cause 2 species of fish to
be extinct."
Story Workbench Semiautomaticpractical
NLU using QRG-CEcontrolled language
Human Results
Simulation Results
8
The EA natural language system
Sentence interpretation
Parsing
Task and domain specific reasoning
Discourse interpretation
Word-sense disambiguation
Quantifier scoping
Sentence attachment
Semantic role assignment
QP frame construction
Situation reification
Anaphora resolution
Intra-sentential anaphora resolution
Temporal ordering
QP frame and process rules
Semantic frames
QRG-CEgrammar
COMLEX Lexicon
KB

• Novel combination of off-the-shelf components
• ResearchCyc KB contents (1.2M facts)
• Comlex lexicon
• Allens parser
• DRT-based semantic interpreter
• Originally developed by Kuehne (2004) for
  modeling roles of qualitative representations in
  NL semantics

9
Cultural Models of Food Webs

• How groups conceive of relationships in the
  natural world
• Experiments carried out by Medins group
• Participants given scenario about a perturbation
  to a population.
• Asked to predict effects
• Example
• E Do you think that the disappearance in the
  bears would affect other plants and animals in
  the forest?
• P Well, there probably be a lot more berry
  growth for example, because they wouldnt be
  eating the berries. There probably would be a lot
  less maybe dead trees, because they wont be
  climbing on the trees and shredding them.

10
Qualitative Concept Maps

• QCM provides a scientist-friendly interface for
  encoding causal models using Qualitative Process
  (QP) Theory (Forbus, 1984)
• QCM uses a concept map interface (Novak
  Gowin,1984)
• QCM automatically checks for modeling errors,
  provides detailed feedback.

11
Bears Disappearing Example
12
Qualitative Concept Maps
there probably be a lot more berries because they
wouldnt be eating the berries
Number of bears influence the eating rate
Bears eat berries
Number of berries increase as eating rate
decreases
13
Experiment Detecting culture via models

• Experiment Automatic classification of the
  models based on the cultural group they belonged
  to
• Data Interviews collected by Medins group

Menominee0
E.A.0
Menominee1
SEQL
SEQL
Menominee Generalization
European American Generalization
E.A.1
Menominee2
E.A.2
Menominee3
E.A.3
Example
Construct generalizations from concept maps
Measure similarity of new model to
generalizations about cultures
SME
SME
Example X is more similar to culture Y
14
Results Culture classification via analogical
generalization
i.e., if a test model for Bears Disappearing was
from Menomonee, the system correctly categorized
it 65 of the time

• 81 models encoded using QCM, in response to 3
  food web scenarios
• Results are averaged over 1000 trials. Trial 4
  models from each group chosen randomly for
  generalization, 8-10 models randomly chosen for
  test set
• Conjecture for improving accuracy Increase
  uniformity in follow-up questions during
  interviews.

15
Can Inspect Generalizations for Insights

• The number of facts that were consistent across
  individuals was higher in Menominee models
• The number of consistent causal relations was
  higher among Menominee
• On average, there were 24 facts found
  consistently across all Menominee models vs. 16
  facts for non- Menominee
• On average, Menominee models contained 4 causal
  relations whereas non-Menominee models only
  contained 2.

16
Experiment Classification via expertise

• What can we learn from automatic classification
  of the models based on the level of expertise?
• Hunters and fishermen are considered experts
  within this domain
• Analogical processing results
• Classifying experts from non-experts within
  Menomonee models 72.5.
• Classifying experts from non-experts within
  European American models 52 (almost chance)
• Suggests Menomonee are more influenced by their
  daily activities than European Americans
• Consistent with independent manual analysis
• Menominee hunters more likely than Menominee
  non-hunters to mention ecological relations (19.8
  vs 10.14, p lt 0.01)
• No significant difference between ecological
  relations mentioned by hunters versus non-hunters
  for European-Americans (16.08 vs 16.22, p lt 0.97)

17
Next Steps

• Possible source of noise Degree of follow-up
  questioning varied in interviews
• Working with Medins group to figure out
  practical protocols to get more uniform data
• Closing the loop Making predictions from
  automatically generated models
• The same participants could be re-interviewed, or
  more individuals from same group, depending on
  level of modeling
• Use initial interviews for gathering training set
• Construct generalizations, make predictions
• Conduct more interviews to test predictions

18
Modeling Moral Decision-Making

• Goal Model effects found in literature on moral
  dilemmas (e.g., Trolley Problem)
• Sacred values vs secular values
• Quantity sensitivity
• Differences in group responses
• Given a scenario S, outcomes A, B that depend
  on what action is taken
• Predict which action someone would prefer

??
19
MoralDM model
Rules for extracting relevant quantities,
producing valuations
Utility Calculator
Decision
New Dilemma
SME
Cultural or life stories
Vary according to group
Prior Cases w/Decisions
Protected Values
20
Protected Values and Quantity Insensitivity

• Protected values (PVs) concern acts and not
  outcomes
• People with PVs show insensitivity to quantity of
  outcomes (Baron and Spranca 1997, Lim and Baron
  1997)
• In trade-off situations, they are less sensitive
  to the consequences of their choices
• Quantity insensitivity of PVs are not absolute
  (Bartels and Medin 2007)

21
Modeling Protected Values

• Idea Use order of magnitude formalism from QR
  to model protected versus secular values
• Introduces stratification in values
• Degree of stratification can be varied to model
  context effects

22
Order of Magnitude

• Based on Dague (1988) formalization
• A K B ? A-B K Max(A,B)
• A and B are equivalent
• A !K B ? A-B gt K Max(A,B)
• A and B are comparable magnitudes tell which is
  greater
• A K B ? A lt K B
• B dominates A, or A negligible w.r.t. B
• K determines how stratified values are
• K can be adjusted to account for different
  sensitivities towards consequences
• K 1/10 20 gt 15
• K 1/3 20 K 15
• K 2 20 gtgt K 15

23
An Example Dilemma

• A convoy of food trucks is on its way to a
  refugee camp during a famine in Africa.
  (Airplanes cannot be used.) You find that a
  second camp has even more refugees. If you tell
  the convoy to go to the second camp instead of
  the first, you will save 1000 people from death,
  but 100 people in the first camp will die as a
  result.
• Would you send the convoy to the second camp?
• What is the largest number of deaths in the first
  camp at which you would send the convoy to the
  second camp?

24
Scenarios

• 12 moral decision making scenarios from Ritov and
  Baron (1999) were chosen as inputs
• civil rights, nature preservers, combating
  traffic accidents, Jewish settlements, Arab
  villages,
• Manually translated into predicate calculus
• Goal Semi-automatically translate with EA NLU
• Recent Progress The river scenario was
  automatically translated from simplified English

25
Simplification Example

• Original text
• "As a result of a dam on a river, 20 species of
  fish are threatened with extinction. By opening
  the dam for a month each year, you can save these
  species, but 2 species downstream will become
  extinct because of the changing water level."
• Simplified text
• "Because of a dam on a river, 20 species of fish
  will be extinct. You can save them by opening the
  dam. The opening would cause 2 species of fish to
  be extinct."

26
Example of EA NLU output

• "Because of a dam on a river, 20 species of fish
  will be extinct."
• (explains-Generic
• (thereExists dam44262
• (thereExists river44314
• (and (on-UnderspecifiedSurface dam44262
  river44314)
• (isa river44314 River)
• (isa dam44262 Dam))))
• (thereExists set-of-species44351
• (and (isa set-of-species44351 Set-Mathematical)
• (cardinality set-of-species44351 20)
• (forAll species44351
• (implies (elementOf species44351
  set-of-species44351)
• (and (isa species44351
  BiologicalSpecies)
• (generalizes
  species44351 Fish)
• (thereExists
  extinction44579
• (and (isa
  extinction44579 Extinction)
• 
  (objectActedOn extinction44579
  species44351))))))))
  )

27
Results

• Out of the 12 scenarios, MoralDM makes decisions
  matching those of participants on 11 scenarios
• In 8 scenarios, both first-principles reasoning
  and analogical reasoning provide the correct
  answer
• In 3 scenarios, first-principles reasoning fails,
  but analogical reasoning provides the correct
  answer
• In 1 scenario, both reasoning strategies fail

28
Next steps MoralDM

• Test on wider range of examples
• Scale up story libraries for different cultural
  groups
• Incorporate MAC /FAC for retrieval
• Currently using fables, folktales as sources
• Extend EA NLU and QRG-CE coverage to handle
  range of both cultural stories and interview
  stories
• Story Workbench EA NLU MITs GUI

29
Future Work

• Use automatically constructed cultural models to
  make novel predictions
• Experiment with two-phase interview structure
• Continue extending EA NLU to broader coverage
• Needed to scale automatic model construction
• Extend automatic cultural model construction to
  other kinds of data
• Fables, folk-tales, life stories, valuation
  rules i.e., the culturally-specific inputs to
  MoralDM.
• Test MoralDM model on wider range of problems and
  inputs from different groups
• Data is the limiting factor right now

30
Details
31
Elements of QP Theory

• Physical Process
• All causal changes stem from physical processes.
• Example heat flow between a brick and a room
• Parts of physical processes
• Participants
• Entities participating in a physical process
• Example the brick, the room
• Conditions
• Determine when a process is active
• Example difference in temperature
• Consequences
• Hold as long as a process is active
• Direct influences (derivatives)
• Indirect influences (functional relations)

32
Modeling Blame Assignment

• Context Computational version of attribution
  theory from psychology being developed at ICT by
  Gratch and Mao
• Assigns credit/blame for a consequence C of an
  action A to an agent P based on
• Did P cause A?
• Did P intend C?
• Did P foresee that C would follow from A?
• Was P coerced by another actor?
• Uses simple axioms to assign binary values of
  credit/blame to agents based on
• causal knowledge, expressed by plans
• Simple axioms relating cause, intention, and
  knowledge
• Rules for inferring knowledge and intent from
  dialogue acts

33
Mao Gratchs Computational Model of blame
assignment

• Based on Shavers theory of moral responsibility
  (1985)
• Attribution along dimensions of responsibility
• Judgment of responsibility follows
• Responsibility may lead to blame
• Observed behaviors in a simulation environment
• Plan library, using Hierarchical Task Networks,
  for causal inference
• Speech acts covering order negotiation for
  dialogue inference
• Attribution variables as Boolean assignments
• Infers which agent in the scenario is to blame

34
QR Model (Tomai Forbus, 2007)

• Same causal/dialogue input, different attribution
  process
• Qualitative representation provides more rigorous
  modeling method
• Social science theories describe dimensions of
  responsibility are described as continuous
  parameters
• Predictions, experimental results cast as ordinal
  relationships
• Qualitative modeling captures this directly,
  without ad hoc step of constructing quantitative
  equations and postulating numerical parameters
• Also fits the data better

Tomai, E., and Forbus, K. 2007. Plenty of Blame
to Go AroundA Qualitative Approach to
Attribution of Moral ResponsibilityProceedings
of QR-07
35
Whos to blame?

• The chairman of Beta Corporation is discussing a
  new program with the vice president of the
  corporation.
• The vice president says, The new program will
  help us increase profits, but according to our
  investigation report, it will also harm the
  environment.
• The chairman answers, I only want to make as
  much profit as I can. Start the new program!
• The vice president says, Ok, and executes the
  new program.
• The environment is harmed by the new program.

(From Mao 2006, adapted from Knobe 2003)
36
Modes of Judgment
Foreseen?
Intended?
Coerced?
No
No
No
Yes
Unforeseen
Unintended
Voluntary
Coerced
Increasing responsibility

• Four distinct modes of judgment
• Responsibility is strictly increasing
• Translates to six views
• Within each mode responsibility is qualitatively
  proportional to an attribution variable

37
Maos Results
38
Survey Results
VP 1 3.73
Chair 2 5.63
Chair 1 3.00
Chair 4 4.13
VP 4 5.20
VP 3 3.23
lt
lt
lt
lt
lt
VP 2 3.77
Chair 3 5.63
39
QR Model Results
VP 4 5.20
Chair 1 3.00
Chair 4 4.13
Chair 2 5.63
VP 2 3.77
VP 3 3.23
lt
lt
lt
lt
VP 1 3.73
Chair 3 5.63
40
QR Model Results
Unforeseen
Coerced
Voluntary
VP 4 5.20
Chair 1 3.00
Chair 4 4.13
Chair 2 5.63
VP 2 3.77
VP 3 3.23
lt
lt
lt
lt
VP 1 3.73
Chair 3 5.63
41
QR Model Results
Unforeseen
Voluntary
Coerced
VP 4 5.20
Chair 1 3.00
Chair 4 4.13
Chair 2 5.63
VP 2 3.77
VP 3 3.23
lt
lt
lt
lt
VP 1 3.73
Chair 3 5.63

• Violates strict ordering of modes of judgment
• Challenges an assumption of Shavers theory